Binucleate cell recognition in automated gynecologic cytopathology

Medium resolution two dimensional image analysis techniques were applied to a large number of abnormal and binucleate cells from the uterine cervix. Techniques were developed for the extraction of nuclear boundary information from fluorescence images and two new nuclear shape descriptors were applied to the binucleate cell recognition problem. The descriptors were a contextual measure of concavity location and size on the nuclear boundary, and mean area of the two largest concavities. These features were superior both to total convex deficiency and to ratio of perimeter squared to area when applied to binucleate cell recognition. Classification based on the second feature, mean convex deficiency, provided lowest error rates. These were approximately 10% false positive and false negative single cells, and occurred with an effective two micron spot size. Higher false negative rates were observed with a one micron spot size.